Information-Aware Decision Making in Teams of Autonomous Vehicles and Humans
Abstract
The goal of this project is to develop methods for optimizing sensing andcommunication decisions in heterogeneous teams of autonomous vehicles and humans to allowfor efficient querying of high-impact information. Teamwork requires consulting with each otherfor" assistance. In the future, this assistance may come from autonomous vehicles, missioncommanders, and humans deployed in the field."" However, determining when to ask forassistance and what type of assistance is needed (e.g., information, guidance, or clarificatio"n) is achallenge. Even human teams struggle with asking for help. One of the major lessons taught inschool is how to work independently and how to ask for guidance when needed. Hierarchicalcommand structures in management and military are designed to alleviate" these issues, but thesestructures are often rigid and require constant communication. Such communication is typicallyunavailable"" for autonomous vehicles in the field. In future naval missions, a key requirement willbe to adjust decisions with changing mission"" objectives using only intermittent, low-bandwidthcommunication. To enable this critical capability, the PI will develop new decisi""on makingalgorithms that are closely tied to mission objectives, ranging from target identification, tracking,and neutralization t""o mapping, classification, and scene understanding.Technical Approach: Classical information theoretic measures (e.g., mutual infor""mation) arefundamentally incapable of capturing complex dependencies in information gathering problems.However, recent advances in" machine learning theory have developed measures with broaderapplicability based on reducing the edge weight in an undirected graph". Instead of optimizing thevalue of information, these measures optimize the discriminating power of observations relativeto possi""ble actions. However, such measures are not currently generalizable to distributedheterogeneous systems. To fill this gap, the PI w""ill (1) develop theoretical foundations andalgorithms for autonomously trading off between sensing, communication, and querying for""human assistance in autonomous vehicle teams using measures of the discriminating power ofobservations, and (2) provide implementa"tion and validation of this approach in simulation andexperiments on autonomous vehicles.Anticipated Outcome of Research: The algorithms and theoretical analysis developed in thisproject will directly advance decision making for autonomous vehicles operating wi"th humans.By shifting the paradigm from value of information to discriminating power of observations, thisresearch will allow for" heterogeneous teams of vehicles and humans to perform tasks withcomplex relationships between sensing and action. These contributi"ons will benefit the robotics,artificial intelligence, and machine learning communities by generating the fundamental theoryand al""gorithms for fully distributed decision making in teams of aerial, marine, and groundvehicles performing mapping, tracking, surveil""lance, and reconnaissance tasks.Impact on DoD Capabilities: The proposed techniques will directly advance the Office ofNaval Resea"rch~s thrust towards autonomy and unmanned systems interacting with humans. Theframework for determining the impact of potential observations and active queries will allowmission commanders to interact with teams of heterogeneous autonomous vehicles performingnaval-relevant tasks. The developed algorithms will adjust for different degrees of interactionand remain applicable in environments with highly limited and intermittent communication.These capabilities will support tactical and strategic naval planning and improve the combatreadiness of unmanned naval systems in a variety of environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2017
- Source ID
- N000141712581
Entities
People
- Geoffrey A. Hollinger
Organizations
- Office of Naval Research
- Oregon State University
- United States Navy